Visually-evoked choice behavior driven by distinct population computations with non-sensory neurons in visual cortical areas

It is widely assumed that variability in visual detection performance is attributed to the fidelity of the visual responses in visual cortical areas, which could be modulated by fluctuations of internal states such as vigilance and behavioral history. However, it is not clear what neural ensembles represent such different internal states. Here, we utilized a visual detection task, which distinguishes perceptual states to identical stimuli, while recording neurons simultaneously from the primary visual cortex (V1) and the posterior parietal cortex (PPC). We found distinct population dynamics segregating hit responses from misses despite no clear differences in visual responses. The population-level computation was significantly contributed by heterogenous non-sensory neurons in V1, whereas the contribution from non-neurons with the previous outcome selectivity was prominent in PPC. These results indicate different contributions of non-sensory neurons in V1 and PPC for the population-level computation that enables behavioral responses from visual information.


Introduction
1 Identical sensory stimuli sometimes evoke different perceptual and behavioral responses. For 2 instance, in a sensory detection task, human or animal subjects are instructed or well-trained to 3 reliably report the presence and absence of sensory stimuli to obtain rewards. When the sensory 4 evidence is near the threshold for the decision criterion, the subjects' reports vary across trials 5 despite subjects' best efforts to get rewards. Interestingly, even if they report the absence of 6 stimuli, it is sometimes possible that they could correctly guess the contents of the stimuli above 7 chance level if they are forced to answer 1-4 . Revealing the neural mechanisms underlying such 8 trial-by-trial variability of perceptual reports is crucial to understand how the brain exploits 9 sensory information for optimal decision making.

11
The trial-by-trial variance of the responses to identical stimuli is believed to reflect noises in 12 sensory information conversion into motor outputs 5 . It has been demonstrated that the 13 variability of firing rates of sensory neurons is responsible for the trial-by-trial variability of 14 choices 6,7 . However, the accumulating evidence suggests that perceptual decision is also 15 significantly affected by latent subjective states reflecting task engagement 8 . For instance,

41
Together, these studies support the notion that V1 and PPC form distinct cortical states at a 42 population level that integrates task-relevant external signals 42 with internal states for subjective 43 detection performance.

45
Though neural imaging studies addressed population coding of sensory processing across 46 different cortical areas, the go/no-go task paradigm, which is often used in those experiments 47 with head-fixed animals, is susceptible to subjective biases: go trials may contain false alarms, 48 and no-go trials may contain false rejections 43 due to fluctuating internal states as described 49 above. To further classify such internal states during the visual detection task, we previously 50 developed a spatial-visual cue detection task for free-moving rats 44 . The task combines a two-51 alternative spatial choice task with a third option for void stimulus, which allowed us to 52 differentiate three distinct visual detection states based on choice types during the task. By

Rats performed visual detection task based on their internal threshold 67
We trained seven rats to perform a spatial visual-cue detection task (Fig.1a), which is 68 essentially a 3-alternative choice design and encourages animals to report the presence and the 69 absence of the peripheral visual stimuli as described in our previous study 44 . Briefly, the rat

83
also showed >90% correct rejection performance when the visual stimulus was not emitted in 84 3C trials (Fig.1d, Extended Data Fig.1). These results confirmed that rats have a generalized 85 strategy to make peripheral choices across stimuli only when their internal detection criterion is 86 met. Also, rats showed correct peripheral choices above the chance level when the shutter 87 forced them to make peripheral choices after choosing to stay in the central port (Fig.1e, gray, 88 Extended Data Fig.2). Thus, rats received visual information but did not exploit it maximumly.

89
We labeled trials with different choice types as Hit, Miss+, and Miss-according to the choice 90 performances (Fig.1f). To find out the influences of trial history in those choice types, we 91 conducted a generalized linear model (GLM) analysis for spatial choice (left and right) and 92 hit/miss choice (Fig.1g-h). In FC trials, about half of the choice variance was explained by the 93 Figure 1 | Neural recording in V1 and PPC during spatial visual-cue detection task. a, Schematics of behavioral paradigm. Rats initiated a trial by nose-poking into the central port and waited for 0.2-0.6s to receive a peripheral stimulus. Rats were rewarded by poking to the corresponding peripheral port when the peripheral stimulus was presented. Rats were rewarded in the central port when the peripheral stimulus was not presented. FC and 3C trials were identical except that the central port was shut in FC trials when rats kept nose-poking in the central port more than 0.5s. b, Spatial choice accuracy in 3C and FC trials with graded visual contrast in session A and a fixed contrast in session B. c, Miss rate in 3C trials. d, Correct rejection rate in 3C trials. e, Spatial choice accuracy before and after shutter closure in FC trials. f, Conceptual illustration of three distinct choice types. We classified spontaneous correct response, forced correct response and forced error choice as Hit, Miss+ and Miss-, respectively. g-h, The impact of task parameters on behavioral variability using GLM fitting with several task parameters to Left/Right choice in 3C and FC (g), and to Hit/Miss choice in peripheral stimulus trials (h). i, Illustration of kernel regression approach to fit the neural activity by three task predictors. j, Example trial-averaged responses (plot with the shaded area) and model predictors (Thick line). k, Fraction of stimulus preferring and non-preferring neurons in V1 (left) and PPC (right). l-n, Trial-averaged neural activity for each choice type in Enhanced type stimulus preferring neurons, Suppressed type stimulus preferring neurons, and stimulus non-preferring neurons in V1 and PPC. All responses were z-scored, and neurons were sorted by max peak latency in Hit trials.
model showing significant influences from previous peripheral rewards and the current stimulus 94 direction. In contrast, in 3C trials, influences of rewards disappeared, and 77% of the choice 95 variance was accounted by stimulus alone (Fig.1g). In addition, a relatively small portion of the 96 total variance (10%) for hit/miss choices was explained with the model containing past rewards, 97 in which the major contribution was previous reward ipsilateral to the current stimulus (Fig.1h).

98
Thus, our task segregated trial-by-trial variability of responses during the detection task into

139
We found that significant separation between choice types at the analysis window in both V1

V1 Noise correlation was increased in forced detection performance before and 236 after stimulus presentation 237
Our results thus far demonstrate that both stimulus preferring and non-preferring neurons are 238 essential for choice types in V1 and PPC. However, we used "pseudo-population" that 239 combined neural activity recorded in different trials in these analyses. Therefore, our analysis 240 missed considering the correlation structure of pairs of simultaneously recorded neurons within 241 each trial (i.e., noise correlation). If the noise is closer to random across neurons (low noise 242 correlation), information coding is more reliable and efficient 46 . We first examined classification 243 accuracy in each session using a simultaneously recorded population. Only the V1 population 244 showed significantly higher classification than shuffled data (Fig.5a, p = 0.0460, p = 0.1985 in 245 V1 and PPC, respectively). Next, we compared classification accuracy with a decorrelated 246 population in which each neuron in the same session was randomly taken from different trials.

247
Only the V1 population was significantly decreased classification accuracy in the decorrelated  (Fig.5b), indicating that the correlation structure, at least, in V1 was crucial for 249 population computation.

251
To investigate the correlation structure in V1 and PPC for choice types, we calculated noise 252 correlations in 200ms sliding time windows relative to stimulus onset (Fig.5c, -0.4 -0.4s from 253 stimulus onset). We found that noise correlation in Miss+ trials increased in -0.2s -0.2s from 254 stimulus onset compared to Hit in V1 neuron pairs, while PPC neuron pairs did not differ in 255 choice types (Fig.5d). Such a difference in the noise correlation was only apparent in neuron 256 pairs between regular-spiking neurons (Extended Data Fig.6d). To examine whether reduced 257 noise correlation is associated with previous outcomes, we compared the noise correlation between previous rewarded and unrewarded trials. We found that noise correlation in V1 was 259 significantly reduced in previous ipsilateral reward trials around stimulus presentation timing 260 (Fig.5e), while that was significantly increased when animals were previously rewarded in the

278
Instead, we showed multiple evidence for population-level computation that contributes to the 279 correct responses to visual stimuli (Hit/Miss+) in V1 and PPC. First, despite no apparent 280 differences in averaged activity among choice types (Fig.1l-n), we found a specific divergence 281 between choice types at the multiple levels of the population activity, particularly with the 282 robust contribution from the stimulus non-preferring neurons in both V1 and PPC (Fig.2).

283
Second, during pre-stimulus and stimulus epochs, contributions of individual sensory non-284 preferring neurons for classification of choice types were significantly greater compared to the 285 stimulus preferring neurons in PPC but not in V1 (Fig.3g). Third, we found that individual 286 neurons that poorly represented choice types also contributed population coding of choice types 287 especially in stimulus presentation timing (Fig.3a-f). Fourth, these contributions were more 288 stable in stimulus non-preferring population in PPC but not in V1 (Fig.4). Finally, V1 neuron 289 pairs, but not PPC, showed increased noise correlation in Miss+ trials before and during visual 290 stimulus presentation (Fig.5c-d), indicating that the V1 population gains efficiency of 291 information processing for future behavioral output.

292
It has been postulated that stochastic behavioral responses to identical sensory stimuli are 294 generated by fluctuations of background neural ensembles preceding to the external inputs 47 .

295
Our model-free analysis revealed that there are at least three distinct population dynamics that 296 differentiate behavior to identical stimuli (Fig.2)

318
Therefore, we conclude that the value-based decisions do not solely explain the recruitments of 319 non-sensory neurons in V1 and PPC. On the other hand, it is possible that irreverent movements 320 during the task performance could have affected Hit responses due to suboptimal head/body 321 positions 49,50 . However, our data shows the noise-correlation in V1 is specifically increased in 322 Miss+ trials, suggesting that, at least, the accurate performance in Hit trials is due to an intrinsic 323 population-level mechanism that can be related to sensory-motor transformation during the

438
Visual cue detection task 439 The visual cue detection task design was described previously 2 . The task was comprised of 440 randomly interleaved three choice (3C) and forced-choice (FC) trials with equal probabilities in 441 a session. The only difference between the trial types was that, in FC trials, the central port was 442 shut with the shutter door to prevent the rat from continuing to central nose poke (Fig. 1a). After  nose withdrawal from the central port was also treated as miss error, though it occurred rarely 460 (<5%). There was no punishment in any error trials and the next trial was allowed to be initiated 461 after ITI. In the no-signal trials, animals need to wait for 0.2-0.6 s without stimulus and another 462 0.5-1 s to get a reward from the central port. There was no cue to distinguish the initial delay 463 (0.2-0.6 s) and reward delay (0.5-1 s). Thus, animals did not have any external clue to 464 differentiate the signal trials from the no-signal trials, except for the presentation of the signal 465 itself. In the FC trials, the shutter was closed 0.5 s after stimulus presentation onset, and the rats 466 were forced to choose either the left or right port (Fig.1a right). In cases where no stimuli were 467 presented in FC trials, the animals were never rewarded.

468
Session B was the same protocol with session A described above except that only single 469 stimulus difficulty was used. We applied the stimulus contrast level of 70-85% accuracy in FC 470 trials in session A.

Surgery
Rats were anesthetized with 2.5% isoflurane before surgery, and it was maintained throughout 474 surgical procedures. We monitored body movements and hind leg reflex and adjusted the depth 475 of the anesthesia as needed. An eye ointment was used to keep the eyes moistened throughout 476 the surgery. Subcutaneous scalp injection of a lidocaine 1% solution provided local anesthesia 477 before the incision. A craniotomy was performed over the anterior part of the right V1 (AP -478 6.36 to -7.32 mm, ML 3.2 mm relative to the bregma, 0.2 to 0.4 mm below the brain surface) 479 and right PPC (AP -3.8 mm, ML 2.5 mm relative to the bregma, 0.2 to 0.4 mm below the brain 480 surface) and a custom-designed electrode was vertically implanted using a stereotactic

510
Behavioral data analysis 511 Spatial choice accuracy was the percentage of the correct peripheral port choices in trials where 512 either peripheral port was chosen upon the presentation of the peripheral stimulus (Fig.1b). The 513 miss rate was the percentage of central port choices in trials where visual stimuli were presented 514 in 3C trials (Fig.1c). The correct rejection rate was the percentage of central choices in trials 515 where visual stimuli were not emitted in 3C trials (Fig.1d). Reaction time was defined as the    To identify visual responsive neurons, we used a time-locked kernel regression approach 547 ( Fig.1i, j, Extended Data Fig.4a-c). In this approach, the firing rate of recorded neurons is 548 described as a linear sum of task predictors aligned to task events. In this study, we considered 549 the stimulus onset and reaction timing kernels. According to this kernel, the predicted firing rate

560
To determine whether each neuron is sensitive to visual response, we prepared a predictor 561 matrix with full kernels (real design matrix) and matrix of which target kernel is shuffled within

565
This procedure using a shuffled design matrix was repeated 1000 times for statistical 566 measurement. If the F-statistic of the real design matrix scored high value compared to the 95% 567 percentile of the shuffled design matrix, the neuron was deemed selective for the target kernel 568 (Extended Data Fig.4c). In the case of neurons that were selective to the contra-stimulus kernel, or Hit trials, and a negative value indicated to suppress (Extended Data Fig.5a, b)

592
For visualization (Fig.1l-n) and analysis, firing rates were z-scored relative to trial-by-trial 593 baseline rates (from the window -0.4 to 0s).

596
We evaluated the statistical significance in the analysis using data resampling with a 597 bootstrapping procedure 51 . We estimated the P value for the bootstrapping procedure by  subspace defined as the square root of ′ 2 ( ) (Fig.2d, g, i), as follows.

640
For classification by adding-in subsets of neurons (Fig.3b, d), the whole population was

644
For classification of the simultaneously-recoded population ( Fig.4a-b), we first extracted 645 sessions with at least five neurons in each region and at least 20 trials in each condition (20 646 sessions). We trained the classifier as the same procedure described above and predicted the test 647 data. In the decorrelated population in V1 and PPC (Fig.4b), the trials were shuffled within trials in each neuron. We then calculated the classification accuracy of real data and the 649 decorrelated population as the same procedure described above.

650
To measure the contributions of each neuron for the choice types (Fig.3g)

670
To estimate the stability of population coding, we calculated Pearson's correlation coefficients 671 between neuronal weights of each classifier at time and (Fig.4c). To quantify 672 time-resolved decay of population activity pattern, we used Pearson's correlation coefficients in 673 0 -0.2s from time (Fig.4e). For comparison between populations, we used t-test for 674 each time point in V1 and PPC (Fig.4f, top), and one-way ANOVA followed by post-hoc Tukey 675 tests for subpopulation (Fig.4f, below).

679
The noise correlation was defined as the correlation coefficients between the noise of neuron 680 pairs to a given visual stimulus using z-scored firing rate for a given visual stimulus. We applied